Advancing No-Reference Point Cloud Quality Assessment via Leveraging Prior Knowledge From Ranking Learning

Abstract

Point clouds have emerged as a crucial representation for 3D visual content, driving innovations in immersive media applications and broadcasting. The development of no-reference point cloud quality assessment (PCQA) methods is crucial for automating quality monitoring and enhancing user experiences. However, the advancement of PCQA is hindered by the limited scale and diversity of existing databases, coupled with the insufficient exploitation of the intrinsic data information and relationships. To overcome these limitations, we first introduce the PCL-PCQA database, a newly constructed large-scale PCQA database comprising 1,599 subjectively evaluated distorted point cloud samples with diverse geometric and textural distortions, derived from 39 high-quality reference point clouds and labeled with reliable mean opinion scores (MOS) following the ITU-R BT.500 standard. Second, we propose a novel ranking-based no-reference PCQA model, Rank-PCQA, which leverages prior knowledge from inherent quality rankings embedded in distortion levels through a two-stage learning pipeline. This approach involves a pre-training stage with pairwise ranking loss and level-wise boundary regularization to learn hierarchical quality relationships, followed by fine-tuning for accurate MOS regression. Extensive experiments on the SJTU-PCQA, WPC, LS-PCQA, and our PCL-PCQA databases demonstrate that our Rank-PCQA achieves state-of-the-art or highly competitive performance against existing advanced methods, validating its effectiveness and strong generalization capability. The database and source code will be available at https://github.com/KunChun/Rank-PCQA